Why Signal Quality Decides Whether AI Can Hear a Child’s Heart Murmur Correctly

A heart murmur in a child is one of the most common reasons for a referral to pediatric cardiology, and it is also one of the most ambiguous. Roughly two thirds of children will have a murmur at some point during childhood, yet between 30 and 75 percent of murmur referrals are eventually classified as innocent, meaning the sound reflects normal blood flow rather than disease. That gap between what a clinician hears and what it actually means has long driven the search for tools that can sort harmless murmurs from dangerous ones automatically. A new commentary published in Pediatric Research by Lizelle v van Wyk and Jacomine Grobler of Stellenbosch University argues that the success of such tools may hinge less on the sophistication of the machine learning classifiers themselves and more on a quieter, more technical question: how clean is the heart sound signal that reaches the algorithm in the first place?

The commentary centers on a recent study by Malvermi and colleagues, who tackled a component of automated auscultation that has received far less attention than classifier design. Digital stethoscope studies in children have proliferated over the past decade, employing artificial neural networks, support vector machines and convolutional neural networks to distinguish normal, innocent and pathological murmurs. Reported sensitivities across these studies range from 84 to 100 percent, with specificities between 75 and 99 percent. Those figures suggest that detection is largely a solved problem. The real clinical challenge, the authors argue, is discrimination, and that challenge begins with the physics of the recording itself.

Heart sound recordings in children are notoriously vulnerable to acoustic contamination. Respiratory sounds bleed into the cardiac signal, respiratory arrhythmia alters the timing of heartbeats, patient motion introduces noise, and the mismatch between a stethoscope bell designed for adults and the smaller chest of a child degrades the capture of sound. Conditions that clinicians take for granted in adult practice simply do not translate to pediatric auscultation. Studies of remote diagnosis have also shown that agreement between clinicians interpreting recorded lung and cardiac sounds can be poor, which means that even human experts struggle when the signal is compromised. Signal denoising, the commentary stresses, is therefore not an optional refinement but a prerequisite for accurate telemedicine.

The denoising approach used by Malvermi and colleagues relies on a technique called non-negative matrix factorisation, which decomposes a recording into a small number of recurring sound sources, each of which can only add to the signal rather than cancel it. In adults, this works reasonably well because pulmonary and cardiac sounds occupy distinguishable parts of the acoustic space. In children, however, the two overlap too closely for direct separation. The researchers’ solution was to shuffle the order of segmented heartbeats before applying the factorisation. Because respiratory noise and motion artifacts are not locked to the cardiac cycle, shuffling destroys their coherence, while anything genuinely tied to the heartbeat survives. What remains is a cleaner heart sound, from which poor-quality recordings can be excluded before classification even begins.

Once the signal was cleaned, support vector machines were applied to the systolic and diastolic phases of each heart sound at each of the five standard auscultatory sites. The reported results were a specificity of 81.8 percent, a sensitivity of 66.7 percent and an accuracy of 77.4 percent, with a composite rule requiring agreement across all sites. Two design choices, however, complicate direct comparison with other studies. Low-quality recordings were excluded after the data had already been split into training and test sets, meaning test sets differed in size and composition between auscultatory sites. The composite rule also counted a diagnosis as correct only when all ten support vector classifiers agreed with the expert assessment, a far more demanding standard than the any-positive-site aggregation used elsewhere, and borderline cases were counted as errors. Both conventions push the reported figures downward relative to studies that do not apply them.

The most consequential decision in the study, according to the commentary, was the exclusion of innocent murmurs altogether. These murmurs proved intermittent and low in amplitude, and they are known to vary with posture and haemodynamic state, making reliable labelling difficult. But removing them changes the task entirely. A referring clinician must distinguish innocent from pathological murmurs, not simply detect whether a murmur exists. Zhou and colleagues quantified that difference on a dataset of 742 recordings from 138 children, supplemented by the CirCor DigiScope collection. Their deep learning models achieved an area under the curve of 0.88 to 0.94 for distinguishing murmur present from absent, but only 0.66 to 0.75 for distinguishing innocent from pathological. A specificity figure from a cohort that excludes innocent murmurs, the commentary warns, cannot be interpreted as an estimate of how many referrals such a system would prevent.

The broader research landscape reinforces this point. Kocharian and colleagues evaluated an automated auscultation device in 563 Iranian children aged two to twelve and reported 97 percent specificity, 90 percent sensitivity and 93 percent accuracy against echocardiography, explicitly motivated by referral from family physicians in rural and underserved areas. Papunen and colleagues, who retained innocent murmurs and interpreted their data against echocardiography, reported 97 percent specificity and 69 percent sensitivity overall, rising to 86 percent sensitivity in children with an audible murmur. The pattern that emerges across these studies is consistent: resource-rich settings stand to benefit chiefly from reduced over-referral and overuse of echocardiography, whereas settings with limited access to subspecialists or echocardiography benefit chiefly from detecting pathological murmurs that would otherwise be missed.

Remote human interpretation offers a useful benchmark for what is achievable. Dahl and colleagues transmitted recordings from 47 children, including no murmur, innocent and pathological cases, to pediatric cardiologists who spent an average of 2.1 minutes per case and achieved a mean sensitivity of 89.7 percent and specificity of 98.2 percent, with inter- and intra-observer kappa values of 0.81 and 0.87. Finley and colleagues sent eight-second recordings from two to four chest locations in 55 children aged one month to 19 years to three blinded pediatric cardiologists, reporting sensitivity of 87 to 100 percent and specificity of 82 to 88 percent. Mahnke and colleagues used a store-and-forward method to classify 103 datasets with 85 percent accuracy, 82 percent sensitivity and 86 percent specificity, rising to 91 and 88 percent when reviewer uncertainty, clinical diagnosis and electrocardiographic abnormalities were factored in.

Field data from low-resource settings add further nuance. Pyles and colleagues collected phonocardiograms from 149 of 7,993 children screened in rural China and transmitted them to a cloud server for asynchronous review by a pediatric cardiologist in the United States. Echocardiography identified 14 children with congenital heart disease, of whom 11 were correctly identified, yielding an accuracy of 91 percent, sensitivity of 78.5 percent and specificity of 92.6 percent. All three missed cases were atrial septal defects misclassified as a venous hum or a Still’s murmur, a confusion that Zhou and colleagues also noted as a persistent overlap between innocent murmurs and this particular defect. The lesson, the commentary argues, is that denoising and classification must be developed together on a dataset that contains innocent murmurs, so that the filtering can be shown to preserve the very sounds the system must eventually distinguish for clinical use.

For remote workflows, automated signal quality assessment has a practical implication that goes beyond classification accuracy. An immediate decision about whether a recording is adequate would permit re-acquisition during the same clinical encounter, sparing families a return visit for re-examination. Both Malvermi and Zhou obtained their recordings under quiet, controlled conditions, so field performance in noisy environments remains unestablished. Behere and colleagues, comparing Eko Core stethoscope recordings with in-person auscultation across normal, innocent and pathological murmurs, reported high agreement with bedside and echocardiographic findings but only moderate inter-rater reliability. Krones and colleagues have further noted that supporting infrastructure and explainability to frontline health workers may be limiting factors in low-income settings. The questions that follow are clear: whether denoising preserves the acoustic features of innocent murmurs, whether performance holds under realistic acquisition, and whether the differences between auscultatory sites can be reproduced in larger test sets. The clinically meaningful endpoint, which none of the studies discussed have yet addressed, is whether such a system actually changes referral behaviour. What the current work does establish is that signal quality is measurable, that it can be improved by exploiting the periodicity of cardiac sound, and that classification performance in children is sensitive to that improvement.

Subject of Research: Signal quality and machine learning classification in pediatric heart murmur auscultation for tele-consultation

Article Title: Pediatric murmur auscultation and tele-consultation: importance of signal quality and classification

Article References: Pediatric murmur auscultation and tele-consultation: importance of signal quality and classification. (n.d.). https://doi.org/10.1038/s41390-026-05537-9

Image Credits: AI Generated

DOI: 10.1038/s41390-026-05537-9

Keywords: pediatric cardiology, heart murmurs, digital stethoscope, signal denoising, machine learning, telemedicine, non-negative matrix factorisation, innocent murmurs, congenital heart disease, auscultation, support vector machines, low-resource settings